Enhancing geographical origin classification of Vietnamese oranges via hybrid preprocessing and supervised learning of UV–Vis spectral data

Due to the high economic value of specific botanical varieties, origin fraud in oranges is one of the top most fraudulent food categories globally, commonly encountered through intentional mislabeling or misrepresentation. These fraudulent behaviors cause global losses of tens of billions of dollars annually and create critical intricacies in quality control, quality assurance, and product reliability. This study proposes a cost-effective, rapid, and non-destructive ultraviolet-visible (UV-Vis) spectroscopy method, accompanied by supervised learning models, to authenticate the geographical origins of Vinh oranges, one of Vietnam’s premium sweet orange cultivars, collected from three key cultivation regions in northern Vietnam (Hoa Binh, Quang Ninh, and Tuyen Quang). To boost classification performance, we applied two preprocessing pipelines: standard normal variate and Savitzky-Golay smoothing (SNV+SG) and in combination with the first derivative (SNV+SG+FD). These were implemented prior to classification using supervised chemometrics (LDA), machine learning methods (random forest, RF), and deep learning models (artificial neural networks, ANN). Non-targeted LC-Orbitrap MS identified key markers, including monosaccharides (glucose, fructose), disaccharides (maltose, Hex-Hex derivatives), organic acids (physodic, glutamic, malate/citramalate family), and flavonoids (hesperidin, luteolin, naringin, kaempferol). Quantitative HPLC-DAD analysis confirmed significant concentration variances relative to geographical origin, which correlated with the electronic transitions (π→π* and n→π*) observed in the 250–1000 nm VIS-NIR spectral range. Using minimal preprocessing, the overall discrimination efficiency of supervised models was remarkable (>90%) in the broadband range, while RF and ANN achieved acceptable accuracy (>80%) in the narrowband (250-400 nm). However, LDA yielded strong classification performance irrespective of the spectrum region and preprocessing level. Notably, the recognition ability of RF and ANN outperformed that of LDA on full-band spectra when data were further derived to first order. These findings emphasize the feasibility of this robust method for practical applications in orange origin classification and quality monitoring.